Projects
Content Agent: AI Research & Publishing System
A LangGraph system that researches and drafts technical articles, checks claims against retrieved evidence, and routes content and layout through human review. Improved exact evidence recall from 0.788 to 0.939 across 33 evaluation queries. Supervised demo + engineering case study.
Code-Fix Agent: Self-Correcting AI System
A goal-directed AI agent that takes broken Python scripts, executes them, diagnoses the failure, patches the code, and retries on its own. Built with a LangGraph state machine, human-in-the-loop approval, LangSmith tracing, and a 95% fix rate across 20 diverse error types.
Knowledge Agent: Persistent RAG with Hybrid Search
Local memory-augmented agent with hybrid retrieval (BM25 + dense + cross-encoder reranking), claim verification, and persistent session memory. 92% accuracy, 100% tool-routing accuracy on adversarial test cases.
CLI Research Agent: Raw Agent Loop + LangGraph Rebuild
Terminal-based research agent built from scratch using OpenAI-compatible tool-calling, with no frameworks involved. Takes a question, searches the web, reads source pages, writes a structured markdown report. Later rebuilt with LangGraph to compare raw loop vs. state machine execution.
Bosch Production Line: Predictive Quality Control
End-to-end ML pipeline predicting manufacturing failures on 1.18M rows with 171:1 class imbalance. Engineered path features revealing 72× failure rate signal: certain station paths fail at 41.7% vs 0.58% global mean. Chunk-aware CV and phased feature roadmap progressing from MCC 0.19 → 0.33, targeting ≥ 0.52.
Silent Recalls: Live Vehicle Safety Monitoring
Production-grade ETL pipeline monitoring NHTSA complaints with live risk tracking. Automated detection of vehicles with dangerous complaint-to-recall ratios. GMC Sierra 1500: 445 complaints, zero recalls. Weekly automated runs with hash-based alerting.
Bearing Failure Prediction: 2.88h Accuracy
Production-grade ML system predicting bearing RUL with 2.88-hour accuracy in critical zones. 10x improvement through weighted loss optimization. Modeled ~$300K annual savings and modeled 98.5% failure prevention in critical zones.
About
I build agentic LLM systems and the evaluation infrastructure that makes them reliable: claim-level grounding verification, hybrid retrieval, reflection loops, and human-in-the-loop control, with observability and cost/latency tracking on every run.
My flagship, Content Agent, runs a full research-to-publish pipeline: it retrieves sources, drafts, verifies every factual claim against those sources, self-critiques, routes through human approval, and publishes live. Every run is measured against a golden evaluation set.
Underneath the agent work is a deeper foundation in production ML (predictive maintenance, manufacturing defect modeling, and automotive safety analytics), built across years in mechanical engineering and at Honda R&D Americas. That domain background is why I treat agent design as a reliability problem first: anticipate the failure modes, measure them, engineer around them.
I'm interested in systems that anticipate failures, learn from context, and operate reliably in production.
Skills
Agentic AI & LLMs
- LLM Tool Calling & Agent Loops
- RAG: Hybrid Retrieval & Reranking
- LangGraph State Machines
- Claim Verification & Grounding
Evaluation & Observability
- Evaluation Harnesses & Golden Datasets
- LangSmith Tracing
- Cost & Latency Tracking
- Failure-Mode & Regression Testing
Machine Learning & Data
- Predictive Modeling & Risk Systems
- Time Series Analysis
- Feature Engineering
- Statistical Analysis
Engineering & Infrastructure
- Python, SQL, PostgreSQL
- FastAPI, Docker, AWS
- ETL Pipeline Design
- GitHub Actions, CI & Netlify
Learning Log
Supervised Learning Models: Technical Notes
A breakdown of the model families you reach for most in production: linear models, tree-based ensembles, and gradient boosting. Deep dives on each.
LightGBM: A Practitioner's Guide
How LightGBM actually works: leaf-wise growth, histogram binning, and why it often beats XGBoost on tabular data. Includes practical tuning notes and common pitfalls.
Data Preprocessing: Technical Notes
A systematic series on preparing data for machine learning: feature scaling, encoding, missing data, and outlier treatment.
Data Preprocessing: Feature Scaling Deep Dive
Why scaling matters, when it doesn't, and how to pick the right method for your model. A practical guide covering StandardScaler, MinMaxScaler, RobustScaler, and the assumptions each one makes about your data.
Contact
Open to AI Engineer roles: full-time, contract, and remote (US / Global).